Digitcity

레이더 센싱 데이터를 이용한 classification 문제에 머신러닝을 이용하고자 하는데, amplitude와 phase 두 attribute의 feature를 1D CNN을 이용하여 각각 추출한 후 모델의 끝에서 다시 병합하여 사용하고자 한다.

Keras에서는 이와 같은 방식의 모델을 구성할 수 있도록 도와주는 functional.api가 존재한다.

The Functional API (keras.io)

위 사이트를 참고하여 아래 사진과 같은 모델을 구성해보았다.

def VGG_branch(X, Y, test_X, test_Y, cp_filepath, EPOCH=100):
    
    amp_input = keras.Input(shape=(X.shape[2],1), name="amplitude")
    phs_input = keras.Input(shape=(X.shape[2],1), name='phase')
    
    amp_features = layers.Conv1D(64, (3), activation = 'relu', input_shape = (X.shape[2], 1), padding = 'same')(amp_input)
    amp_features = layers.Conv1D(64, (3), activation = 'relu', padding = 'same')(amp_features)
    amp_features = layers.MaxPool1D(2)(amp_features)
    amp_features = layers.Conv1D(128, (3), activation = 'relu', padding = 'same')(amp_features)
    amp_features = layers.Conv1D(128, (3), activation = 'relu', padding = 'same')(amp_features)
    amp_features = layers.MaxPool1D(2)(amp_features)
    amp_features = layers.Conv1D(256,(3), activation = 'relu', padding = 'same')(amp_features)
    amp_features = layers.Conv1D(256,(3), activation = 'relu', padding = 'same')(amp_features)
    amp_features = layers.Conv1D(256,(3), activation = 'relu', padding = 'same')(amp_features)
    amp_features = layers.Conv1D(256,(3), activation = 'relu', padding = 'same')(amp_features)
    amp_features = layers.MaxPool1D(2)(amp_features)
    print(amp_features.shape)
    amp_features = layers.Flatten()(amp_features)

    phs_features = layers.Conv1D(64, (3), activation = 'relu', input_shape = (X.shape[2], 1), padding = 'same')(phs_input)
    phs_features = layers.Conv1D(64, (3), activation = 'relu', padding = 'same')(phs_features)
    phs_features = layers.MaxPool1D(2)(phs_features)
    phs_features = layers.Conv1D(128, (3), activation = 'relu', padding = 'same')(phs_features)
    phs_features = layers.Conv1D(128, (3), activation = 'relu', padding = 'same')(phs_features)
    phs_features = layers.MaxPool1D(2)(phs_features)
    phs_features = layers.Conv1D(256,(3), activation = 'relu', padding = 'same')(phs_features)
    phs_features = layers.Conv1D(256,(3), activation = 'relu', padding = 'same')(phs_features)
    phs_features = layers.Conv1D(256,(3), activation = 'relu', padding = 'same')(phs_features)
    phs_features = layers.Conv1D(256,(3), activation = 'relu', padding = 'same')(phs_features)
    phs_features = layers.MaxPool1D(2)(phs_features)
    print(phs_features.shape)
    phs_features = layers.Flatten()(phs_features)

    x = layers.concatenate([amp_features, phs_features], axis = -1)
    x = layers.Dense(4096, activation = 'relu')(x)
    x = layers.Dense(4096, activation = 'relu')(x)

    material_output = layers.Dense(4, activation = 'softmax', name = 'material_output')(x)

    model = keras.Model(inputs = [amp_input, phs_input],
                        outputs = [material_output],)

    model.summary()
    keras.utils.plot_model(model, "./multi_input_and_output_model.png", show_shapes=True)
    model.compile(optimizer = 'adam', loss = tf.keras.losses.CategoricalCrossentropy(), metrics=['accuracy'])
    checkpoint_filepath = cp_filepath
    callback = tf.keras.callbacks.ModelCheckpoint(
        filepath = './model/' + checkpoint_filepath,
        monitor='val_accuracy',
        mode='max',
        save_best_only = True,
        save_weigths_only = False,
    )

    model.fit(
    {"amplitude": X[:,0,:], "phase": X[:,1,:]},
    {"material_output": Y},
    epochs=EPOCH,
    validation_data = ({"amplitude" : test_X[:,0,:], "phase": test_X[:,1,:]}, {"material_output" : test_Y}),
    callbacks = [callback]
)